Google open-sources RRSI: agents self-improve their harness, lifting Terminal-Bench 2.1 to 80.2%
sudoraohacker · x · 2026-09-29
Google Research released RRSI (Regularized Recursive Self-Improvement), which lets an agent automatically evolve its own harness—prompts, control flow, tools, memory and context management—while keeping the model frozen, using evaluation to decide which edits to keep.
To counter harness-evolution overfitting (big in-distribution gains that vanish out of distribution), RRSI regularizes both sides:
- Proposal side: an annealed budget caps bundled edits; the proposer is conditioned on the full edit history so falsified hypotheses aren't redrawn; stalled runs get redirected to unexercised components
- Selection side: a critic screens candidates for suite-specific logic before evaluation; gains must exceed evaluation noise; stale components get pruned
Per the project report, with Claude Opus 4.8 fixed, Terminal-Bench 2.1 rose from 74.2% to 80.2%, and SWE-bench Verified (not used for selection) rose from 82.0% to 83.8%. Paper and project page are linked; real-world gains still need on-task validation.
Related event: Google Open-Sources RRSI for Recursive Agent Self-Improvement(2 posts)→
More from coding & agent
- Personal agents from big labs vs one shared company agent: which design wins? — AshishKulkarni1411 · 2026-11-16
- WebArena-Pro launches: 300-task multimodal web-agent benchmark across 20 apps — xhluca · 2026-11-16
- Self-improving agent papers this week: agents learn the benchmark, not the task — TheTuringPost · 2026-11-16
- 1,000 agents, one week, 630k lines: HF builds a weights-free Python chatbot — _lewtun · 2026-11-16
- One-person content agency targets $30K/mo with 11 AI agents — full playbook — VibeMarketer_ · 2026-11-16
- Hugging Face releases definitive fine-tuning guide for 3D generation model TRELLIS.2 — ariG23498 · 2026-11-16